Reduce network load for cache-sync of authorization-data

[danielFesenmeyer]

We currently use Akka Distributed Data for synchronizing a authorization-cache (thing-cache, policy-cache) between all services which either read (search, gateway) or update (things, policies) this data. This works fine, but creates unnecessary network load because the whole cache is distributed to all shards, even if a shard does not need the data. Another disadvantage of the current approach is that the "Akka Distributed Data" cache cannot be cleared, which may cause memory problems in the long run.

In this task, we should at least address the following issues:

• remove policy-cache from search: sync via events and full-sync fallback (without cache) has been tested and works fine
• replace "Akka Distributed Data" with a local cache per shard, holding only the data relevant for this shard
  • implement the cache as an LRU cache with a configurable maximum size
  • the cache-entries should be lazy-loaded on first access
  • update the cache whenever the cached data changes (e.g. by Events via PubSub), the things-service and policy-service should no longer write to the cache actively
  • to make sure that the cache is eventually consistent even when events get lost, let the cache entries expire after a configurable time: a re-load will be necessary on next access
  • to further optimize memory usage, remove entries from the cache which have not been accessed for a configurable time

Caffeine seems to be a good choice for a local cache implementation: https://github.com/ben-manes/caffeine

[danielFesenmeyer]

Implemented by PR https://github.com/eclipse/ditto/pull/163.